Mining Goes Digital – Mueller et al. (Eds)
© 2019 Taylor & Francis Group, London, ISBN 978-0-367-33604-2
101
Machine learning classification of geochemical and
geophysical data
L. Huang
The University of Sydney, Sydney, NSW, Australia
M. Balamurali & K.L. Silversides
Australian Centre for Field Robotics, The University of Sydney, Sydney, NSW, Australia
ABSTRACT: Accurate classifications in downhole exploration data are essential for exploration, geological modelling and prediction of mining outputs. These holes are typically interpreted manually, which is slow, subjective and prone to error. Machine learning techniques
can potentially automate these classifications. Our test deposit contains the shale dominated
West Angelas (WA) Member, and banded iron formation and iron ore in the Mount Newman (MN) Member. Deep learning using autoencoders, Support Vector Machine (SVM)
and k-means were applied to classify stratigraphy and rock type using mineral groups or
geochemical assays. Autoencoder produced accuracies of 83.2–95.1%, K-means accuracies of 30.4–58.8%, and SVM accuracies of 81.6–88.5%. Geochemical assays were a better
indicator of stratigraphy. Autoencoder and SVM produced significantly better results than
k-means. This was probably due to k-means not using training data. Although they gave
similar results, autoencoder is preferable to SVM for this application as it can handle more
than two categories.
1 INTRODUCTION
Accurate classifications in downhole exploration data are essential for exploration, geological modelling and for prediction of mining outputs such as ore tonnages. The data is typically
divided into groups based on properties such as stratigraphy and ore or waste designations.
While interpreting data from these holes is typically done manually, the process is slow and
subjective and is prone to error. Different machine learning techniques can potentially automate some of these classifications. Many different types of machine learning exist, and
methods such as Gaussian Processes (Silversides & Melkumyan 2016), t-SNE (Balamurali &
Melkumyan 2016) and continuous profile model (Nathan et al. 2017) have been applied to
deposits that are similar to our test case. One way of grouping these methods is by the way
that the model is trained, for example supervised or unsupervised learning.
This study aimed to investigate the usefulness of different types of machine learning methods to identify geological features from exploration hole data. Three classification methods
were used: deep learning using autoencoders, support vector machine (SVM) and k-means.
These methods were done as semi-supervised training, supervised training and clustering
without training data respectively.
Our test case is a banded iron formation (BIF) hosted iron ore deposit in the Mamma
Mamba Iron Formation in the Hammersley Region of Western Australia. This deposit contains two significant stratigraphic units. These are the West Angelas (WA) Member, which is
shale dominated, and the Mount Newman (MN) Member, which is a BIF that is mineralised
to iron ore in some locations (Clout 2006; Lascelles 2000). The rock can be generally divided
into three classes, ore BIF and shale. In this study we use different machine learning methods
© 2019 Taylor & Francis Group, London, ISBN 978-0-367-33604-2
101
Machine learning classification of geochemical and
geophysical data
L. Huang
The University of Sydney, Sydney, NSW, Australia
M. Balamurali & K.L. Silversides
Australian Centre for Field Robotics, The University of Sydney, Sydney, NSW, Australia
ABSTRACT: Accurate classifications in downhole exploration data are essential for exploration, geological modelling and prediction of mining outputs. These holes are typically interpreted manually, which is slow, subjective and prone to error. Machine learning techniques
can potentially automate these classifications. Our test deposit contains the shale dominated
West Angelas (WA) Member, and banded iron formation and iron ore in the Mount Newman (MN) Member. Deep learning using autoencoders, Support Vector Machine (SVM)
and k-means were applied to classify stratigraphy and rock type using mineral groups or
geochemical assays. Autoencoder produced accuracies of 83.2–95.1%, K-means accuracies of 30.4–58.8%, and SVM accuracies of 81.6–88.5%. Geochemical assays were a better
indicator of stratigraphy. Autoencoder and SVM produced significantly better results than
k-means. This was probably due to k-means not using training data. Although they gave
similar results, autoencoder is preferable to SVM for this application as it can handle more
than two categories.
1 INTRODUCTION
Accurate classifications in downhole exploration data are essential for exploration, geological modelling and for prediction of mining outputs such as ore tonnages. The data is typically
divided into groups based on properties such as stratigraphy and ore or waste designations.
While interpreting data from these holes is typically done manually, the process is slow and
subjective and is prone to error. Different machine learning techniques can potentially automate some of these classifications. Many different types of machine learning exist, and
methods such as Gaussian Processes (Silversides & Melkumyan 2016), t-SNE (Balamurali &
Melkumyan 2016) and continuous profile model (Nathan et al. 2017) have been applied to
deposits that are similar to our test case. One way of grouping these methods is by the way
that the model is trained, for example supervised or unsupervised learning.
This study aimed to investigate the usefulness of different types of machine learning methods to identify geological features from exploration hole data. Three classification methods
were used: deep learning using autoencoders, support vector machine (SVM) and k-means.
These methods were done as semi-supervised training, supervised training and clustering
without training data respectively.
Our test case is a banded iron formation (BIF) hosted iron ore deposit in the Mamma
Mamba Iron Formation in the Hammersley Region of Western Australia. This deposit contains two significant stratigraphic units. These are the West Angelas (WA) Member, which is
shale dominated, and the Mount Newman (MN) Member, which is a BIF that is mineralised
to iron ore in some locations (Clout 2006; Lascelles 2000). The rock can be generally divided
into three classes, ore BIF and shale. In this study we use different machine learning methods